MULTIMODALITY MEDICAL IMAGE FUSION USING BLOCK BASED INTUITIONISTIC FUZZY SETS

Transcription

1 ISSN: Soundrapandiyan et al. ARTICLE OPEN ACCESS MULTIMODALITY MEDICAL IMAGE USION USING BLOCK BASED INTUITIONISTIC UZZY SETS Rajkumar Soundrapandiyan*, Rishin Haldar, Swarnalatha Purushotham and Arvind Pillai School of Computing Science and Engineering, VIT University, Vellore, INDIA ABSTRACT Image fusion combines more than one image from various environments into a single image. This can be useful for subsequent processing of the image, especially in medical imaging where it can help in disease diagnosis. This paper uses the block based Intuitionistic uzzy Sets() to fuse the multimodality medical images. s can effectively handle the inherent uncertainties of digital images. Initially, in this model, entropy is used to deduce the optimal parameter value for defining the membership and non-membership function. This, in turn generates the Intuitionistic uzzy Images (II) from the original image. inally, the IIs are partitioned into image blocks and then recombined by the generated membership function. This paper compares the proposed method with popular ones like Principal Component Analysis (PCA), simple averaging (AVG), Laplacian Pyramid Approach(LPA), Discrete Wavelet Transform (DWT) and MPA (Morphological Pyramid Approach) on various performance measures such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Peak Signal to Noise Ratio (PSNR), Structural Similarity Inde (SSIM), Universal Image Quality Inde (UIQI), Mean and Standard Deviation (STD). The eperimental results show better image visualization generated through the proposed method compared to the other methods, in overall.. Received on: 8 th -Nov-015 Revised on: 8 th -eb-016 Accepted on: 31 st Mach-016 Published on: 19 th May-016 KEY WORDS Medical image fusion; Intuitionistic fuzzy image; Entropy; Quantitative measures; Multimodality images. *Corresponding author: rajkumars@vit.ac.in INTRODUCTION DNA microarray Image fusion is widely used as an effective technique for analysis of images [1]. These images are obtained from various domains like satellite images, biometrics, robotics, remote sensing etc., and there are customised image sensors for each of these domains. Consequently, the data obtained from these specialised sensors may be incompatible with each other. or eample, in medical imaging, the image generated by an MRI machine gives clear details of soft tissues while a CT (X-Ray) machine gives clear details of bone structures. In this scenario, if we are required to find the clear details of both, or more, of the features, where the data is incompatible, image fusion can be an effective tool to address the issue. This gives us the motivation to apply image fusion on medical images. Image fusion can be carried out by mainly two techniques, spatial fusion and transform fusion. Based on the unification phases, fusion can be done in three levels, namely piel, feature and decision levels. Piel level fusion combines the piel values directly and creates a composite image. The simplest method just takes the average of the piel values of source images. Laplacian pyramids [], PCA [3] are some of the other techniques which use piel value fusion. In order to improve upon the degraded performance of the average policy of fusion algorithm, many multi-resolution transform techniques emerged, like pyramid decomposition, wavelet transforms [4] etc. usion of the images by singular value decomposition (SVD) [5] works quite well on piel basis and outperforms PCA. The MSVD [6] technique, which looks into multiple properties like sphericity, isotropy and self-similarity of signals, performs faster than SVD. Image fusion technique also generated a highly featured picture using multiscale decomposition. The various attempts in using multi-scale transform showed that shifting of invariance is highly desirable for image fusion. In this contet, NSCT [7], a complete transform, has been very effectively utilized in image fusion. Guest Editors Profs. Swarnalatha & Prabu Soundrapandiyan et al. 016 IIOABJ Vol

2 Image processing, however, has many uncertainties at every phase. uzzy sets [8] have been known to remove these uncertainties, especially in luminance and contrast of the image. In medical images, poor luminance increases the uncertainty of the image, and [9], an improvement on the traditional fuzzy set, has been quite successful in removing these uncertainties. Thus, by using the multimodal properties of the image as well as using fuzzy sets, the image fusion can be etremely effective. One paper [10] uses Intuitioinstic uzzy Sets on multimodal images to fuse the images, and the results were very encouraging. This paper presents a new way to fuse more than one medical image, and builds on the efforts done [10]. This paper also uses the block based Intuitionistic uzzy Sets () to fuse the multimodal medical images. However, a new and customised entropy function is used to deduce the optimal parameter value for defining both the membership and the non-membership function. This generates the Intuitionistic uzzy Images (II) from the original images. inally, the IIs are partitioned into image blocks and then recombined by our generated membership function. The reconstructed fused image has high degree of luminance and contrast. The resultant pictures are provided for subjective evaluation. This paper also objectively compares the proposed method with popular methods like simple averaging (AVG) [11], Principal Component Analysis (PCA) [3], Laplacian Pyramid Approach(LPA) [], Discrete Wavelet Transform (DWT) [1] and MPA (Morphological Pyramid Approach) [13] by various performance measures such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Peak Signal to Noise Ratio (PSNR), Structural Similarity Inde (SSIM), Universal Image Quality Inde (UIQI), Mean and Standard Deviation (STD). The eperimental results are very encouraging and show that the proposed method, overall, has performed much better than these popular methods. The following sections give the specific details of our work. Section II describes our proposed methodology along with the required computational models. Section III describes the performance measures through which we are evaluating our proposed technique. Section IV describes the eperimental results and its subjective and objective comparison with the other popular methods. inally, we conclude in Section V. PROPOSED METHODOLOGY The block diagram of our proposed method is shown in ig. 1. The individual steps carried out in our proposed method, are as follows: ig: 1. Block diagram of the proposed method.. 1. Read/Accept the input images. There are si datasets of images, each of size 5656 piels.. uzzification of input images using Equation (1). 3. Generation of intuitionistic fuzzy image using Equation (8). 4. Divide the image into blocks of size use the each block based on the value of entropy using Equation (9). 6. Defuzzification of the fused image using Equation (10). uzzification uzzification [14] is the first step of fuzzy image processing. It consists of converting the image from spatial domain into the fuzzy domain. It can be defined as T 1 ma d e (1) Soundrapandiyan et al. 016 IIOABJ Vol

3 Here ma is the maimum intensity level for the given input image; e and d represent the eponential and denominational fuzzifiers, respectively. When ma = then, µ = 1 indicating the maimum brightness. uzzifier d is calculated using Equation () and e is assigned to constant value. d ma 1 1 min min 1 ().. Intuitionistic uzzy Image (II) In general, piel values of images have ambiguity and uncertainty. However, some uncertainty still remains while specifying the brightness of image piels. The main objective of the proposed method is to remove the ambiguity in those image piels. To address this issue, the image is converted from fuzzy domain to intuitionistic fuzzy domain. The intuitionistic fuzzy domain has an additional property of degree of hesitation compared to fuzzy domain. The hesitation degree is used to align the membership function values within a range. This can effectively remove the uncertain gray level values of ambiguous image piels [10]. An Intuitionistic uzzy Set () is epressed in terms - [19,], by (3) Based on the Equation (1), the degree of the membership function of II is computed as (4) The degree of the non-membership function is computed as The {(, ( ), ( ), ( )) X} ; 11, 0 1 ; 1, 0 1 ; ; ; (5) degree of hesitation is defined as (6) The parameter used. The entropy is defined (7) 1 ENT ; PQ P 1 Q1 i0 j0 ; ; ; ; ; ;, ;, ;, ; 0,1,..., L 1 In Equation (7), the value of corresponds to the highest value of entropy. inally, the II is defined as (8) Entropy based image fusion To fuse the images, the obtained resultant image of 1 and from Equation (8) is decompose into m X n blocks and denote the T th image block of two decomposed images by 1T and T respectively. The entropy based fusion process is defined as fuse 1T min, 1T ma, 1T T T T if ENT ( if ENT ( 1T 1T ) ENT ( ) ENT ( otherwise Soundrapandiyan et al. 016 IIOABJ Vol T T ) ) (9)

4 where ma and min represent the maimum and minimum operations in. Defuzzification Equation (10) epresses the defuzzification process to convert the image from fuzzy domain [3] to the spatial domain where (i,j) is the final fused image. fuse 1 fuse ma d * d 1 e ( i, j) T (10) EVALUATION MEASURES The measurement and analysis on the fused images are done both objective as well as subjective quality measures. This effectively helps in better assessment of the information in the images. or the subjective measure, pictorial representations of the images are provided. or the objective analysis, the following measures are used. In all the measures defined here, R and represent the intensity value of the reference (original) image and the fused image at coordinates i, j respectively and P, Q denote the width and height of the image. Root Mean Square Error (RMSE) It is a [1] method to measure the differences between values predicted by an ideal (reference) image and the fused images. It is calculated as P Q 1 RMSE ( R ) P Q i1 j1 (11) RMSE for the reference and fused images will increase with decrease in similarity, and approaches zero whenever they are similar. Mean Absolute Error (MAE) It is a method which measures the mean of the absolute error between the reference and fused images. P Q 1 MAE R (1) P Q MAE also increases with decrease in similarity between reference and fused images and vice versa. Peak Signal to Noise Ratio (PSNR) PSNR [15] is a method used to measure the quality of the fused image with respect to the reference image. It is defined as: PSNR 10log 10 MAX MSE (13) MSE pq p q R( i, j) ( i, j) i0 j0 i1 j1 where MAX is the maimum value in an image and MSE is the mean square error value of the image. (14) Structural Similarity Inde (SSIM) It provides a way to measure the similarity between the two images. SSIM is an improved version of the peak signal to noise ratio [16]. It is defined as Soundrapandiyan et al. 016 IIOABJ Vol

5 (( (15) R C1) ( R C)) SSIM (( R C1) ( R C)) where µ and µ R denote the average intensities of image and R, σ and σ R denote the variance of image and R, σ R gives the covariance of and R, C 1 and C are constants. The SSIM inde value varies from -1 to 1. When two images are identical, this value will turn out to be 1. Universal Image Quality Inde (UIQI) It is a method to measure the quality of the images [14]. This quantifies the amount of data that has been transferred from the ideal image to the resultant fused image. UIQI defines image distortion by a combination of three factors, namely contrast distortion, loss of correlation and luminance distortion. (4 R )( R ) UIQI (16) ( )( ) where σ R is the covariance of R, µ and µ R denote the average intensities of image and R, σ and σ R denote the variance of image and R. The UIQI inde value varies from -1 to 1. Once again, a 1 indicates the identical nature of the two images. Mean (MEAN) R R The mean intensity estimates the luminance of an image. This is deduced by P Q 1 MEAN PQ i1 j1 Standard Deviation (SD) It shows the etent of variation or dispersion from the average or mean [17,0]. Standard deviation takes into account the original image and the acquired transmission noise. Absence of any noise in the transmitted image increases its effectiveness and portraits the image s contrast. SD can be calculated as (18) P Q 1 SD MEAN P Q i1 j1 (17) RESULTS AND PERORMANCE EVALUATION The eperimental results of the fusion techniques are analyzed with si brain images taken from CT and MRI (T). Each CT image, combined with T, are considered as one set for fusion. This, in turn, totally derives si combinations of input dataset. All images have the same size of 56 * 56 piels, with 56-level gray scale. Subjective evaluation of results igure- gives the subjective comparison of the results from average method, PCA method, Laplacian method, DWT method, MPA method and proposed method. ig.3. evident that the proposed method generated results with good visualization (i.e. high luminance and contrast) than other eisting methods. Performance Evaluation or the objective measures, the measures discussed in the section III are used. The results generated from the proposed method for each of the measures used to quantify the results, are compared with the average method, PCA method, laplacian method, DWT method and MPA method. The comparative analyses of each of the measures are tabulated in [Table-1,, 3 and 4]. The results for the RMSE measure are tabulated in [Table-1]. It is evident from Table 1 that the RMSE is lower for proposed method compared to other five methods, which means that proposed method introduces very less error. Soundrapandiyan et al. 016 IIOABJ Vol

6 Dataset1 Dataset Dataset3 Dataset4 Dataset5 Dataset6 CT image MRI Image Average PCA Laplacian DWT MPA Proposed method ig:. Comparison (subjective) of the fusion results over 6 images.. Table: 1. Comparative analysis of RMSE usion method Dataset1 Dataset Dataset3 Dataset4 Dataset5 Dataset6 Average PCA Soundrapandiyan et al. 016 IIOABJ Vol

7 Laplacian DWT MPA Proposed Table:. Comparative analysis of MAE usion method Dataset1 Dataset Dataset3 Dataset4 Dataset5 Dataset6 Average PCA Laplacian DWT MPA Proposed Table: 3. Comparative analysis of PSNR usion method Dataset1 Dataset Dataset3 Dataset4 Dataset5 Dataset6 Average PCA Laplacian DWT MPA Proposed Table: 4. Comparative analysis of SSIM usion method Dataset1 Dataset Dataset3 Dataset4 Dataset5 Dataset6 Average PCA Laplacian DWT MPA Proposed The results of the MAE measure are shown in [Table-]. It is observed that the proposed method introduced the least error for five of the si datasets. The results of the PSNR measure are shown in [Table-3]. It is apparent from Table 3 that the PSNR value of each and every dataset is superior for the proposed method, indicating a higher image quality. The results of the SSIM measure are shown in [Table-4]. Table 4 clearly shows that the SSIM value of every dataset is closest to 1, compared to the other five methods, indicating the maimum similarity to the original image. Soundrapandiyan et al. 016 IIOABJ Vol

8 ig: 4. Comparative analysis of UIQI.. ig: 5. Comparative analysis of MEAN.. ig: 6. Comparative analysis of SD.. The UIQI values in igure 4 show that the results of the proposed method for every dataset is closest to 1, compared to others, thus indicating the maimum similarity. The results in igure 5 show that the mean value for the proposed method is more than the other approaches, signifying more teture information on the resultant image. The impressive results are also visible in igure-6, which shows the values for the standard deviation measure. CONCLUSION In this paper, image fusion using block based intuitionistic fuzzy sets has been proposed. Since, the entropy provides teture information of an image, the technique of block comparison with the entropy adopted in the paper Soundrapandiyan et al. 016 IIOABJ Vol

9 as well as the adaptive calculation of the necessary parameter for the process sums up its novelty. The eperimental results show that proposed method provides better visualization than average method, PCA method, laplacian method, DWT method and MPA method. In addition, proposed method confers better result compared to the other eisting methods for both the objective and quantitative measures. urthermore, the fused image obtained from proposed method has been found to be more informative and thereby can be used for efficient disease diagnostics. CONLICT O INTEREST Authors declare no conflict of interest. ACKNOWLEDGEMENT None. INANCIAL DISCLOSURE No financial support was received to carry out this project. REERENCES [1] Stathaki, T. [011] Image fusion: algorithms and applications. Academic Press. [] Burt, P. J., & Adelson, E. H. [1983] The Laplacian pyramid as a compact image code. IEEE Transactions on Communications 31: [3] Sun, J., Jiang, Y., & Zeng, S. [005] A study of PCA image fusion techniques on remote sensing. International Conference on Space information Technology: 59853X X. [4] Li, H., Manjunath, B. S., & Mitra, S. K. [1995] Multisensor image fusion using the wavelet transform. Graphical models and image processing 57: [5] Kakarala, R., & Ogunbona, P. O. [001] Signal analysis using a multiresolution form of the singular value decomposition. IEEE Transactions on Image Processing 10 : [6] Lung, S. Y. [00] Multi-resolution form of SVD for tet-independent speaker recognition. Pattern recognition 35 : [7] Da Cunha, A. L., Zhou, J., & Do, M. N. [006] The nonsubsampled contourlet transform: theory, design, and applications. IEEE Transactions on Image Processing 15: [8] Ross, T. J. [009] uzzy logic with engineering applications. John Wiley & Sons. [9] Atanassov, K. T. [1986] Intuitionistic fuzzy sets. uzzy sets and Systems 0 : [10] Balasubramaniam, P., & Ananthi, V. P. [014] Image fusion using intuitionistic fuzzy sets. Information usion 0: [11] Sharmila, K., Rajkumar, S., & Vayarajan, V. [013] Hybrid method for multimodality medical image fusion using Discrete Wavelet Transform and Entropy concepts with quantitative analysis. International Conference on Communications and Signal Processing (ICCSP): [1] Pu, T., & Ni, G. [000] Contrast-based image fusion using the discrete wavelet transform. Optical Engineering 39: [13] Wang, Z., Ziou, D., Armenakis, C., Li, D., & Li, Q. [005] A comparative analysis of image fusion methods. IEEE Transactions on Geoscience and Remote Sensing 43: [14] Soundrapandiyan, R., & PVSSR, C. M. [015] Perceptual Visualization Enhancement of Infrared Images Using uzzy Sets. In Transactions on Computational Science XXV : [15] Rajkumar, S., & Mouli, P. C. [014] Infrared and Visible Image usion Using Entropy and Neuro-uzzy Concepts. ICT and Critical Infrastructure: Proceedings of the 48th Annual Convention of Computer Society of India- I : [16] Prakash, C., Rajkumar, S., & Mouli, P. V. S. S. R. [01] Medical image fusion based on redundancy DWT and Mamdani type min-sum mean-of-ma techniques with quantitative analysis. International Conference on Recent Advances in Computing and Software Systems: [17] Rajkumar, S., & Kavitha, S. [010] Redundancy Discrete Wavelet Transform and Contourlet Transform for multimodality medical image fusion with quantitative analysis. 3rd International Conference on Emerging Trends in Engineering and Technology : [18] Rajkumar, S., Bardhan, P., Akkireddy, S. K., & Munshi, C. [014] CT and MRI image fusion based on Wavelet Transform and Neuro-uzzy concepts with quantitative analysis. International Conference on Electronics and Communication Systems : 1-6. [19] Szmidt, E. and Kacprzyk, J.[000] Distances between intuitionistic fuzzy sets. uzzy sets and systems 114(3): [0] Gupta, S., Rajkumar, S., Vayarajan, V. & Marimuthu, K.. [013] Quantitative Analysis of various Image usion techniques based on various metrics using different Multimodality Medical Images. International Journal of Engineering and Technology : [1] Purushotham, S. and Tripathy, B.K.. [015] A Comparative Analysis of Depth Computation of Leukaemia Images using a Refined Bit Plane and Uncertainty Based Clustering Techniques. Cybernetics and Information Technologies 15(1): [] Purushotham, S. and Tripathy, B. [014] A comparative study of RICM with other related algorithms Soundrapandiyan et al. 016 IIOABJ Vol

10 from their suitability in analysis of satellite images using other supporting techniques. Kybernetes 43(1): [3] Bhargava, R., Tripathy, B.K., Tripathy, A., Dhull, R., Verma, E. and Swarnalatha, P. [ 013] Rough intuitionistic fuzzy C-means algorithm and a comparative analysis. In Proceedings of the 6th ACM India Computing Convention : ABOUT AUTHORS Mr. Rajkumar Soundrapandiyan is currently working as Assistant Professor (Senior) in School of Computing Science and Engineering, VIT University, Tamil Nadu, India. He received his BE in Computer Science and Engineering and ME in Computer Science and Engineering from Anna University, Chennai, India in 008 and 010 respectively. He is pursuing his PhD at VIT University, Vellore, India. He has published more than ten research papers in reputed international conferences and journals. He is a life member of CSI. His research interest includes digital image processing, computer vision, object detection and target recognition. Mr. Rishin Haldar is currently working as Assistant Professor (Senior) in School of Computing Science and Engineering, VIT University, Tamil Nadu, India. He received his BE in Computer Science and Engineering from Karnatak University, Dharwad and MS in Computer Science from George Mason University, airfa, USA. He is a life member of CSI. His research interests include artificial intelligence, soft computing, bioinformatics and semantic web. Swarnalatha Purushotham is an Associate Professor, in the School of Computer Science and Engineering, VIT University, at Vellore, India. She pursued her Ph.D degree in Image Processing and Intelligent Systems. She has published more than 50 papers in International Journals/International Conference Proceedings/National Conferences. She is having 14+ years of teaching eperiences. She is a member of IACSIT, CSI, ACM, IACSIT, IEEE (WIE), ACEEE. She is an Editorial board member/reviewer of International/ National Journals and Conferences. Her current research interest includes Image Processing, Remote Sensing, Artificial Intelligence and Software Engineering. Mr. Arvind Pillai is pursuing his B.Tech in Computer Science and Engineering in VIT University, Tamil Nadu, India. His research interest includes digital image processing, computer vision and object detection Soundrapandiyan et al. 016 IIOABJ Vol